File size: 7,521 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
// Auto-generated file. Do not edit!
//   Template: src/f16-gavgpool/unipass-neonfp16arith.c.in
//   Generator: tools/xngen
//
// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/gavgpool.h>


void xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c24(
    size_t rows,
    size_t channels,
    const void* input,
    size_t input_stride,
    const void* zero,
    void* output,
    const union xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows != 0);
  assert(rows <= 7);
  assert(channels != 0);

  const uint16_t* i0 = input;
  const uint16_t* i1 = (const uint16_t*) ((uintptr_t) i0 + input_stride);
  if XNN_UNPREDICTABLE(rows < 2) {
    i1 = (const uint16_t*) zero;
  }
  const uint16_t* i2 = (const uint16_t*) ((uintptr_t) i1 + input_stride);
  if XNN_UNPREDICTABLE(rows <= 2) {
    i2 = (const uint16_t*) zero;
  }
  const uint16_t* i3 = (const uint16_t*) ((uintptr_t) i2 + input_stride);
  if XNN_UNPREDICTABLE(rows < 4) {
    i3 = (const uint16_t*) zero;
  }
  const uint16_t* i4 = (const uint16_t*) ((uintptr_t) i3 + input_stride);
  if XNN_UNPREDICTABLE(rows <= 4) {
    i4 = (const uint16_t*) zero;
  }
  const uint16_t* i5 = (const uint16_t*) ((uintptr_t) i4 + input_stride);
  if XNN_UNPREDICTABLE(rows < 6) {
    i5 = (const uint16_t*) zero;
  }
  const uint16_t* i6 = (const uint16_t*) ((uintptr_t) i5 + input_stride);
  if XNN_UNPREDICTABLE(rows <= 6) {
    i6 = (const uint16_t*) zero;
  }

  const float16x8_t vscale = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.scale));
  const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
  const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));
  for (; channels >= 24; channels -= 24) {
    const float16x8_t vi0x01234567 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
    const float16x8_t vi0x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
    const float16x8_t vi0xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
    const float16x8_t vi1x01234567 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
    const float16x8_t vi1x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
    const float16x8_t vi1xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;

    const float16x8_t vi2x01234567 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
    float16x8_t vacc01234567 = vaddq_f16(vi0x01234567, vi1x01234567);
    const float16x8_t vi2x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
    float16x8_t vacc89ABCDEF = vaddq_f16(vi0x89ABCDEF, vi1x89ABCDEF);
    const float16x8_t vi2xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
    float16x8_t vaccGHIJKLMN = vaddq_f16(vi0xGHIJKLMN, vi1xGHIJKLMN);

    const float16x8_t vi3x01234567 = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
    vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
    const float16x8_t vi3x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
    vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi2x89ABCDEF);
    const float16x8_t vi3xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
    vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi2xGHIJKLMN);
    const float16x8_t vi4x01234567 = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
    vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
    const float16x8_t vi4x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
    vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi3x89ABCDEF);
    const float16x8_t vi4xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
    vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi3xGHIJKLMN);
    const float16x8_t vi5x01234567 = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
    vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
    const float16x8_t vi5x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
    vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi4x89ABCDEF);
    const float16x8_t vi5xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
    vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi4xGHIJKLMN);
    const float16x8_t vi6x01234567 = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
    vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
    const float16x8_t vi6x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
    vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi5x89ABCDEF);
    const float16x8_t vi6xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
    vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi5xGHIJKLMN);
    vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);
    vacc89ABCDEF = vaddq_f16(vacc89ABCDEF, vi6x89ABCDEF);
    vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN);

    vacc01234567 = vmulq_f16(vacc01234567, vscale);
    vacc89ABCDEF = vmulq_f16(vacc89ABCDEF, vscale);
    vaccGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale);

    vacc01234567 = vmaxq_f16(vacc01234567, vmin);
    vacc89ABCDEF = vmaxq_f16(vacc89ABCDEF, vmin);
    vaccGHIJKLMN = vmaxq_f16(vaccGHIJKLMN, vmin);

    vacc01234567 = vminq_f16(vacc01234567, vmax);
    vacc89ABCDEF = vminq_f16(vacc89ABCDEF, vmax);
    vaccGHIJKLMN = vminq_f16(vaccGHIJKLMN, vmax);

    vst1q_u16(output, vreinterpretq_u16_f16(vacc01234567)); output = (uint16_t*) output + 8;
    vst1q_u16(output, vreinterpretq_u16_f16(vacc89ABCDEF)); output = (uint16_t*) output + 8;
    vst1q_u16(output, vreinterpretq_u16_f16(vaccGHIJKLMN)); output = (uint16_t*) output + 8;
  }
  if XNN_UNLIKELY(channels != 0) {
    do {
      const float16x8_t vi0x01234567 = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
      const float16x8_t vi1x01234567 = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;

      const float16x8_t vi2x01234567 = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
      float16x8_t vacc01234567 = vaddq_f16(vi0x01234567, vi1x01234567);

      const float16x8_t vi3x01234567 = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
      vacc01234567 = vaddq_f16(vacc01234567, vi2x01234567);
      const float16x8_t vi4x01234567 = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
      vacc01234567 = vaddq_f16(vacc01234567, vi3x01234567);
      const float16x8_t vi5x01234567 = vreinterpretq_f16_u16(vld1q_u16(i5)); i5 += 8;
      vacc01234567 = vaddq_f16(vacc01234567, vi4x01234567);
      const float16x8_t vi6x01234567 = vreinterpretq_f16_u16(vld1q_u16(i6)); i6 += 8;
      vacc01234567 = vaddq_f16(vacc01234567, vi5x01234567);
      vacc01234567 = vaddq_f16(vacc01234567, vi6x01234567);

      vacc01234567 = vmulq_f16(vacc01234567, vscale);
      vacc01234567 = vmaxq_f16(vacc01234567, vmin);
      vacc01234567 = vminq_f16(vacc01234567, vmax);

      if XNN_LIKELY(channels >= 8) {
        vst1q_u16(output, vreinterpretq_u16_f16(vacc01234567)); output = (uint16_t*) output + 8;
        channels -= 8;
      } else {
        float16x4_t vacc0123 = vget_low_f16(vacc01234567);
        if (channels & 4) {
          vst1_u16(output, vreinterpret_u16_f16(vacc0123)); output = (uint16_t*) output + 4;
          vacc0123 = vget_high_f16(vacc01234567);
        }
        if (channels & 2) {
          vst1_lane_u32(output, vreinterpret_u32_f16(vacc0123), 0); output = (uint16_t*) output + 2;
          vacc0123 = vext_f16(vacc0123, vacc0123, 2);
        }
        if (channels & 1) {
          vst1_lane_u16(output, vreinterpret_u16_f16(vacc0123), 0); output = (uint16_t*) output + 1;
        }
        channels = 0;
      }
    } while (channels != 0);
  }
}